원문정보
초록
영어
Regarding to the theories and techniques of cloud computing having been developed and applied in scientific computing field, tasks can be conveniently managed by the cloud platform on the basis of standardized scheduling system with cost (resources consumed) recorded. However, there are two issues which drag the customers’ attention: 1) When will the tasks expect for termination (response time) under a specific resource scheduling; 2) What is the best scheduling solution by considering cost. In order to reply these two questions, a Kriging based forecasting and scheduling system has been proposed in this paper. With the cooperation between the scientific designer and the cloud designer, the design variables for evaluating the cloud applications can be achieved; Kriging surrogate model is then introduced to simulate the approximate functional relationship between the design variables and the response time of the tasks; Sequential quadratic programming optimization algorithm then provides the best scheduling solution for the tasks if cost constraints are to be met. Two real scientific computing cloud applications have been testified on an OpenStack cloud platform, with consequences described in details. The work in this paper has put forward a novel way for the designers and the customers on predictable and reasonable scheduling strategies for the various resource-intensive scientific computing cloud applications with surrogate models and optimization algorithms.
목차
1. Introduction
2. Related work
2.1. Cloud Computing and Task Scheduling
2.2. Application-Level and Resource-Intensive Scheduling Methods
3. Research Background
3.1. Sequential Quadratic Programming
3.2. Kriging Surrogate Model
3.3. Openstack Open-Source IaaS
4. System Design
4.1. Design Variables
4.2. Application-Based “Sampling”
4.3. Kriging Surrogate Model Creation
4.5. Virtual Computing Machine Image and Virtual Computing Application
4.6. Platform Mechanism
5. Testing Cases
5.1. Testing Case One
5.2. Testing Case Two
5.3. Results Analysis and Discussion
6. Conclusion
Acknowledgements
References